Insurance Regulatory Governance
Why Is Our Insurance Company's AI Underwriting Model Biased?
The real regulatory record on AI underwriting bias, what NAIC and state regulators now require insurers to test for, and where the actual proxy-discrimination risk hides.
AI underwriting models usually are not biased by design, they encode bias through proxy variables like ZIP code, credit score, or name-based demographics that correlate with race even without using race directly. NAIC's AI governance bulletin, adopted by 20 plus states, and Colorado's amended testing regulation now require insurers to actively test for exactly this.
cause
How Bias Actually Gets Into an Underwriting Model
Insurance underwriting models almost never use race as an input, and that is precisely why bias is hard to catch. It arrives through proxy variables, ZIP code, credit-based insurance scores, education, occupation, that correlate with race even though none of them mention it. A model trained to minimize loss ratio has no concept that one of its predictive features is a legal liability, it just finds the correlation and uses it.
Documented Precedent: The Allstate Credit-Scoring Case
This is not a hypothetical. Plaintiffs in Texas and Florida sued Allstate in 2001 alleging its credit-based insurance scoring disparately impacted Black and Hispanic customers under the Fair Housing Act. The case settled in 2006: Allstate rolled out a new, public scoring algorithm, added a consumer appeals process, and made payments to the class, without admitting wrongdoing, per Insurance Journal's contemporaneous coverage. It is the clearest real-world example of how a facially neutral scoring variable produced a discriminatory outcome in practice.
The Pattern Is Not Unique to the US
A 2025 academic study in the journal Risks, using 12.4 million quote-bind-claim records from four pan-European life and health insurers, found the lowest-income quintile of policyholders was systematically overcharged, 5.8 percent above the actuarially fair benchmark in life insurance and 7.2 percent in health insurance. This is EU data under the EU AI Act, not a US finding, but the underlying mechanism, an income-correlated proxy driving a pricing distortion nobody explicitly programmed, is the same one US regulators are now requiring insurers to test for.
regulatory
What Regulators Now Require You to Test For
The NAIC's Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted December 4, 2023, requires insurers to maintain a written AI governance program covering risk management, internal audit, and mitigation of adverse consumer outcomes, and gives state regulators explicit authority to request that documentation during an exam. By mid-2026 more than 20 states had adopted it in full or substantially similar form.
New York went further, and earlier: DFS Circular Letter No. 1 (2019) warned that geographic data, credit information, and social media activity 'all have the potential to reflect disguised and illegal race-based underwriting' in life insurance, and asserted the regulator's right to audit an insurer's underwriting algorithms directly. A 2024 follow-up, Circular Letter 2024-7, extended that guidance from life insurance to all lines and to AI systems generally.
Colorado has gone furthest on enforcement mechanics. Its amended Regulation 10-1-1, effective October 15, 2025, extends quantitative bias-testing requirements from life insurance to private passenger auto and health benefit plans, requiring insurers to determine whether their use of external data, algorithms, and predictive models results in unfair discrimination with respect to race. A narrative progress report was due December 1, 2025, with full compliance and annual reporting required by July 1, 2026.
fix
What Actually Reduces Underwriting Bias
The American Academy of Actuaries recommends a practical filter for any variable entering an underwriting model: is it necessary, appropriate, and legitimate for the risk being priced. Applied consistently, that test catches proxy variables like vehicle color or online-only discount eligibility (which quietly disadvantages older customers) before they reach production, not after a regulator finds them.
On the technical side, the same EU study cited above found that adversarial debiasing techniques closed 65 to 82 percent of the premium gap it measured, at a cost of only about 14 basis points to solvency capital, evidence that fixing this is not a binary tradeoff between fairness and financial soundness.
differentiation
How Kriv AI Approaches Insurance Underwriting Model Governance
Kriv AI maps underwriting model governance work directly to the standards above: an AI governance program structured to the NAIC Model Bulletin's requirements, proxy-variable and disparate-impact testing aligned to what Colorado's Regulation 10-1-1 and New York's DFS circular letters now expect regulators to ask for, and documentation an internal audit function can actually use. We do not yet have a named insurance underwriting client case study to publish here, and we will not invent one.
Insurance and other regulated-industry governance engagements start at our $200/hour enterprise floor, with model validation and bias-testing work typically scoped as fractional or advisory engagements in the $300 to $700 per hour range, most engagements beginning at an $8,000 minimum. See our full engagement pricing for the complete breakdown.
Sources
Cited sources
- NAIC: Members Approve Model Bulletin on Use of AI by Insurers (December 4, 2023)
- NY DFS Circular Letter No. 1 (2019): use of external data in underwriting
- Willkie Farr & Gallagher: Colorado Division of Insurance Adopts Amended AI Governance Regulation
- Insurance Journal: Allstate credit-scoring discrimination settlement (June 2, 2006)
- Risks (MDPI): Algorithmic Bias Under the EU AI Act in Life and Health Insurance Underwriting (August 22, 2025)
- American Academy of Actuaries: Unmasking Hidden Bias, Contingencies (Nov/Dec 2025)
Straight answers
Frequently asked questions about Why Is Our Insurance Company's AI Underwriting Model Biased?
Why is our AI underwriting model biased if it does not use race as an input?
Because bias enters through proxy variables, ZIP code, credit score, name-based demographics, that correlate with race even when race itself is never used. A model optimizing for loss ratio has no way to know one of its predictive features is a legal liability.
What does the NAIC AI governance bulletin require insurers to do?
Adopted December 4, 2023 and now in force in 20 plus states, it requires a written AI governance program covering risk management, internal audit, and mitigation of adverse consumer outcomes, with documentation regulators can request during an exam.
Does Colorado now require bias testing for auto and health insurance, not just life?
Yes. Colorado's amended Regulation 10-1-1, effective October 15, 2025, extends quantitative bias-testing requirements to private passenger auto and health benefit plans, with full compliance and annual reporting required by July 1, 2026.
Has an insurer actually been found to discriminate through algorithmic underwriting?
Yes. Allstate's credit-based insurance scoring was the subject of a Fair Housing Act lawsuit alleging disparate impact on Black and Hispanic customers, settled in 2006 with a new public scoring algorithm and a consumer appeals process.
Can debiasing an underwriting model meaningfully close a pricing gap?
A 2025 EU study of 12.4 million insurance records found adversarial debiasing closed 65 to 82 percent of an income-correlated premium gap, at a cost of roughly 14 basis points to solvency capital. That data is EU-specific, not a US finding.
How does Kriv AI help insurers test underwriting models for bias?
We build AI governance programs mapped to the NAIC Model Bulletin and state-specific requirements like Colorado's Regulation 10-1-1, including proxy-variable and disparate-impact testing, priced from our $200/hour regulated-industry floor.
Talk to the team that would do the work
Bring your requirements to a working session with the person who'll actually deliver.
Book a Discovery Call